bible verses, mid-phrase words
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 20 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron lights up on words in the middle of Bible verses and other religious quotes, such as 'all', 'with' or 'enjoy' in the middle of a sentence. Turning it up made the model's writing drift in unrelated directions, so the religious link only shows in the examples.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
God and is profitable for doctrine, for reproof, for correction, for instruction in righteousness: That the man of God may be perfect, thoroughly furnished unto all good works” – 2timothy 3:16-17
-
what I commanded them, saying, 'Obey My voice, and I will be your God, and you will be My people; and you will walk in all the way which I command you, that it may be well with you.'⏎Matthew
-
sparingly shall reap also sparingly; and he which soweth bountifully shall reap also bountifully.⏎Every man according as he purposeth in his heart, so let him give; not grudgingly,<|endoftext|>The good
-
be God, the all-loving creator who meets every human need. In Psalms we read, “Give thanks to the Lord, for he is good; his love endures forever. Let the redeemed of the Lord tell their story. .
-
the Lord for his unfailing love and his wonderful deeds for mankind” (107:1, 2, 8
-
children after them may fear the Lord your God as long as you live by keeping all his decrees and commands that I give you, ad so that you may enjoy log life” (Deuteronomy 6:1, 2). God
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: how to make a one-dimensional system of a one-dimensional **a**-shim and a one-dimensional b-parker-like-2a-arrb-10 - a-see-10 - b-see
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
friend of Jesus who raised him from the dead (John, Chapters 11 & 12).⏎The church survives today.⏎The second mention of
-
born? What did they come to do? (Matthew 2:2)⏎- What prophecy did they chief priests and scribes know about Jesus? (Matthew 2:3-6)⏎- How did Herod scheme against Jesus? How
-
ed the sheep gate--close to the temple. Its name arose either from the sheep market, or from the pool of Bethesda, which was there (Joh 5:2). There the sheep were washed and then taken to the temple for
-
Research Center shows that Muslim countries on average… Latest update on: 2006-SEP-23 Matthew 12:30, Luke 11:23, and Romans 16:17. Unfortunately
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L18:967 suppresses
- L10:446 excites
- L18:3038 suppresses
- L4:2500 excites
- L3:1012 suppresses
- L19:827 suppresses
- L10:934 suppresses
- L15:1274 excites
- L7:247 suppresses
- L9:1141 excites
- L9:2440 suppresses
- L4:1514 suppresses
- L16:1378 suppresses
- L19:175 excites
- L9:2452 suppresses
- L10:315 weakly excites
- L19:1584 weakly excites
- L10:2370 weakly suppresses
- L7:1027 weakly suppresses
- L7:1013 weakly suppresses
- L15:1833 weakly suppresses unclear: mixed mid-sentence words
- L9:131 weakly suppresses
- L9:18 weakly excites text boundary after a document break
- L16:2172 weakly excites
Feeds later layers
- L26:735 excites
- L26:1252 excites
- L21:1886 excites
- L21:2729 weakly excites
- L25:1615 weakly excites
- L25:115 weakly excites
- L21:1281 weakly excites
- L25:2024 weakly excites
- L27:196 weakly excites
- L25:1613 weakly suppresses
- L25:134 weakly excites
- L27:10 weakly excites
- L25:1653 weakly excites
- L24:1417 weakly excites
- L26:2745 weakly excites
- L24:2638 weakly excites
- L25:581 weakly excites
- L25:1169 weakly excites
- L21:2061 weakly excites
- L21:379 weakly excites
- L24:2078 weakly excites
- L22:1182 weakly excites
- L23:415 weakly excites
- L25:828 weakly excites
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- trí
- 美德
- wisdom
- kad
- 岁月
- furt
- 誉
- жд
- grace
- 嬛
Suppresses
- iders
- lexible
- 该项
- ickey
- amped
- 八个
- arker
- ropical
- irical
- 第二个